Marine Navigation & Sensing Giant Captures Major New Market

Industries: Marine Sensing

Use Case: Image Classification, Machine Vision, Object/Defect Identification

Architecture: Knowledge-First AI (AI-CALM)

A global giant in marine navigation and sensing saw a major opportunity in the environmentally friendly, fixed-net fishing market, worth about $2B and growing rapidly. The challenge is to provide the fishing industry with the ability to “see” into deep ocean water and determine the fish types and counts that have migrated into the giant fixed-net networks, with accuracy.

Expert fishermen can readily interpret sonar echograms manually, but it is impossible to build ML models to do the same based on the available data. By applying Aitomatic Knowledge-First AI, engineers at this firm are able to build applications for their customers. Fishermen now get full visibility into when, where, and how much fish hauls of what types are available in their fixed nets at sea.

We knew all along we needed to capture expert fishermen’s knowledge. We just could not find any product to help us do it.

Fixed-Net Service Value, Annual

$100M

Accuracy Improvement

+35%

Less Annotation Effort

-15%

The Problem

Not Enough Data for Machine Learning

Fixed-net fishing is a type of environmentally friendly, passive fishing. It involves predicting the routes of fish runs and waiting for them to come into large arrays of net. Unlike fishing methods that trawl around indiscriminately to catch large fish quantities from a large boat, this method minimizes damage to the environment and helps preserve resources. The fixed-net method is best applicable to catching sardine, horse mackerel, sea bream, yellowtail and squid. The net consists of a network of large hedge nets in the paths of migrating fish, and bag nets to catch them.

The goal is to know when, where, and how large a haul of what types of fish are available in these fixed nets at any given time. This way, fishermen can maximize their profits and minimize disturbance per trip out to their fixed nets.

From the machine-learning perspective, there is very little data available, be it echograms that are sufficiently annotated and labeled, or seasonal history that is geographically dependent. By the time you stratify any dataset to these variables, the dataset obviously becomes too sparse.

Echograms in and around the fixed nets can be produced using modern sonar sensors. But they are impossible to read without experience. On a case-by-case basis, expert fishermen are able to interpret them to assess in real-time the likely number and types of fish present at that particular moment.

So the challenge becomes: how can AI engineers at this marine-sensing company build a high-performing, predictive fixed-net application—without waiting for years worth of catch data to build up?

The Solution

Knowledge-First AI Comes To The Rescue

1

Translate
domain-specific knowledge

2

Combine it with
Machine Learning

3

Deploy Production-ready
ML Solution

For the marine-sensing company, the secret sauce is access to expert knowledge. Even without sensors, an experienced fisherman can tell what fish you’re likely to encounter just based on time of year, time of day, and local weather. They know where and when fish schools will likely appear, and what depths in local waters to find them. They even know which fish will scare off other fish, e.g., no other fish are ever present when yellow-tail are around.

Accuracy has been excellent.
Our customers hauled 45 tons of sardines today!

k-SWE

Knowledge can be applied to different segments

Using the Knowledge-First k-SWE (“Segmented-World Ensemble”) architecture, one first applies all this expert knowledge to create highly conditioned, stratified datasets. These data sets are used to train models specific to each segment. Even rule- or heuristic-based expert knowledge can be incorporated, such as “no other fish will be found near yellow-tail”. The models are then ensembled together to make generalized predictions.

With k-SWE, AI engineers have been able to smoothly combine image- or vision-based modes with heuristic-based modes, hence pooling together the entire knowledge- and data-based resources accessible to the marine-sensing company. As a result, they have been able to get to market quickly with a new, unique, and powerful product offering for their customers.